Godot over MCP: giving an agent something worth calling

MCP standardises how an agent discovers and calls tools. That is genuinely useful and it is also the part that is basically solved. Wiring a server to an engine is an afternoon of work for anyone who has done it before.

What that afternoon gets you, if you stop there, is a set of thin wrappers: create a node, set a property, run a script. An agent can call those all day and still not be able to make a game, because none of them are operations at the level a game is described in. You have given the model a keyboard, not a language.

The interesting question is what the tools should be. That is a design problem about games, not about the protocol.

A tool is worth having when it collapses something the model would otherwise have to invent. "Place terrain of this shape" is a tool. "Set node property" is a keystroke. The first one lets a model be right on the first try; the second one lets it be wrong in a hundred small ways.

The same holds for water, weather, sky, characters and the things they do. Every one of those that becomes a real capability is a body of code the model no longer has to write from nothing, and a class of failure that stops happening.

That is the actual work behind Stage Engine, and it is why it is a capability layer rather than a wrapper. The consequence is one that matters commercially as well as technically: when the layer is good, a cheap model composes reliably against it, and you do not need an expensive one for anything except taste and recovery.

Stage Engine ships with Director 1, which is the agent in the box and writes GDScript into your project from what you describe. But the capability layer is exposed over MCP, so you are not locked to it. If you already work with an agent you trust, it can call the same surface.

Access to that from outside the app is part of the paid plans — free is $0, Indie is $10 a month, Pro is $50. What the paid tiers buy is usage and that external access, not a different engine.

This is bring-your-own-agent rather than bring-your-own-key inside our app. Your agent runs where it already runs, with whatever it already has.

Stage Engine is a desktop application for Mac and Windows with its own copy of Godot inside it. The project it operates on is a real Godot project in a folder on your machine, and when you press play the game runs locally.

Building needs an internet connection because the AI, asset generation and world data all come over the network. Playing what you built does not.

And the division of labour holds even here: the agent writes code, and a person builds the world by hand in the builder. Handing the world-building to a model is not the design, and no amount of protocol changes that.